# Kimi K3 — Moonshot's 2.8T flagship with 1M context lands on web, app, and API > Moonshot AI ships Kimi K3 today, jumping to a 2.8T mixture-of-experts with a 1M-token context and native vision. - Org: Moonshot AI - Date: 2026-07-16 - Category: model - Source: https://platform.kimi.ai/docs/guide/kimi-k3-quickstart - Canonical page: https://ai-tldr.dev/releases/moonshot-kimi-k3/ ## In short Moonshot AI drops the Kimi K3 open weights on Hugging Face, alongside the K3 technical report. K3 is a 2.8T-parameter Mixture-of-Experts model (104B active) with a 1M-token context, Delta Attention hybrid, and native vision — the model is now self-hostable in addition to the kimi.com app and OpenAI-compatible API. ## What it is Kimi K3 is Moonshot AI's new flagship Mixture-of-Experts model, live as of July 16, 2026 on kimi.com, the Kimi mobile app, the Kimi Code coding-agent tier, and the api.moonshot.ai OpenAI-compatible API. K3 takes text, images, and video in, runs with thinking mode always on, and reasons across a 1-million-token context. ## How it works Under the hood the model uses a new Delta Attention hybrid — a linear attention path fused with Attention Residuals — that lets a 2.8T-parameter MoE keep the KV cache tractable at 1M tokens. The API exposes a single `kimi-k3` model id compatible with the OpenAI SDK, streams reasoning and content as separate deltas, supports strict-schema structured JSON, dynamic tool loading, and automatic context caching without a manual cache id. ## Why it matters Kimi K3 is the first Chinese frontier model to cross 2 trillion parameters and ship end-to-end product access on day one. Third-party testers report K3 beating Claude Opus 4.8 and GPT-5.5 on coding benchmarks and landing between GPT-5.6 and Claude Fable 5 overall, which puts Moonshot's price-per-quality within striking distance of Anthropic and OpenAI even before weights or a paper appear. ## Who it's for Engineers who want a 1M-context coding and knowledge model with an OpenAI-compatible API, and teams evaluating Chinese frontier models against Claude and GPT-5. ## Try it curl https://api.moonshot.ai/v1/chat/completions -d '{"model":"kimi-k3","messages":[…]}' ## Quick facts - Maker: Moonshot AI - Size: 2.8T total parameters, 104B active (Mixture-of-Experts, 16 of 896 experts per token) - License: Kimi-K3 License (open weights on Hugging Face) - Context window: 1,048,576 tokens (1M) - Architecture: Kimi Delta Attention (hybrid linear) + Attention Residuals - Modalities in: Text, image, video (native multimodal) - Availability: Open weights on Hugging Face; API kimi-k3 at api.moonshot.ai/v1 (OpenAI-SDK compatible); kimi.com + mobile app + Kimi Code ## FAQ ### How much does Kimi K3 cost to use via API? Moonshot has not published per-token API pricing for Kimi K3 in the quickstart docs yet — the page links out to a chat-k3 pricing table that lists flat pay-as-you-go rates with no context-length tiering. In the Kimi Code product, K3 is gated behind the Moderato plan for 256K context and the Allegretto+ plan for the full 1M-token context. ### How is Kimi K3 different from Kimi K2.7? Kimi K3 jumps from Kimi K2.7's 1T Mixture-of-Experts and 256K context to a 2.8T MoE with a 1M-token context window, and swaps in a new Delta Attention hybrid architecture with Attention Residuals. K3 also adds native visual understanding, whereas K2.7-Code was text-only, and it always runs in thinking mode instead of exposing an instant variant. ### Are Kimi K3 weights open, like Kimi K2.6 was? Moonshot has not published Kimi K3 weights or a Hugging Face model card as of launch. Prior generations — Kimi K2.6 and Kimi K2.7 — were open-weighted on Hugging Face under Modified MIT, but K3 today is only reachable through kimi.com, the Kimi app, Kimi Code, and the Moonshot API. ### How does Kimi K3 compare to Claude Opus 4.8 and GPT-5.5? Independent reviewers report that Kimi K3 outperforms Claude Opus 4.8 and GPT-5.5 on several coding benchmarks and lands between GPT-5.6 and Anthropic's Claude Fable 5 in overall quality. Moonshot has not published its own benchmark table on the K3 quickstart page, so treat the comparisons as third-party until an official model card lands. ### What can Kimi K3 process in a single request? Kimi K3 supports up to 1,048,576 input tokens and up to 1,048,576 completion tokens, with a default max completion of 131,072. Inputs can be text, base64-encoded images, and video files uploaded through Moonshot's files API, and outputs stream as separate reasoning and content deltas so clients can display the thinking trace live. ## Links - [K3 Quickstart Docs](https://platform.kimi.ai/docs/guide/kimi-k3-quickstart) - [Open Weights (HF)](https://huggingface.co/moonshotai/Kimi-K3) - [Tech Report (PDF)](https://github.com/MoonshotAI/Kimi-K3/blob/main/k3_tech_report.pdf) - [Kimi (Web)](https://www.kimi.com/en) - [TechCrunch Coverage](https://techcrunch.com/2026/07/16/moonshots-upcoming-kimi-3-is-expected-to-close-the-gap-with-anthropics-opus-4-8/)